What Are the Qualities of a Great Data Science Manager?
This is something that I’ve been asked many times.
First of all, let me state the obvious: Data science managers need to be good managers in general. Reality is that most technical folks, developers and scientists don’t really care much for managers — or only accept them as a necessary overhead. However as the case has been made by various high profile experiments, the best managers make a huge impact on a team’s performance.
A good manager has a vision, is goal-oriented, cares for the team, listens to them for making decisions, is a mentor and coach, empowers and inspires team members and avoids micromanagement. All software work needs such managers.
On top of this, data science work presents its own sets of unique challenges because it is multidisciplinary, new, iterative, workflow is confusing, results are often hard to interpret and explain, jargon is plentiful, math is tricky for those not familiar with it, some models are a black box and there is hype around what can and can’t be done.
Here’s how Prof. Stein puts it:
Successfully managing a data science team requires skills and philosophies that are different from those that arise in managing other groups of smart professionals. It’s wise to be aware of the potential organizational frictions and trade-offs that can crop up.
So what are the additional desirable characteristics of a great data science (product) manager? I think there are 5 qualities of a highly effective data science manager:
- Balances technical nuances across domains of data, math/stats, machine learning and software and connects them to business context and value
- Earns respect and trust of technical team by contributing at big picture level (are there any blind spots, are we on the right track, challenging assumptions) as well as providing feedback on decisions (eg model selection, level of uncertainty) and interpretation — and importantly,
- Adds structure (e.g. workflow, an agile process with feedback loop and code reviews, code repository, documentation), absorbs shocks and removes barriers, identifies disconnects (say, between business and data science or between data science and dev ops) and builds consensus, facilitates smooth work environment, manages workload, sets the pace and maintains quality
- Takes ownership of key workflow areas such as data acquisition, data quality, prioritizing which aspects are most important, presentation of results etc
- Knows how to launch data science solution for real-world applications .. plans and manages or coordinates the business process changes, production-level code and IT operations needed to bring about the business value
Balázs Kégl talks about “Formalizers”, a special category of data scientists. A slightly edited excerpt from his post is below (square brackets are mine).
They master what is possible at the data science side and what is needed [valuable, feasible] at the business side. They fluently converse with the domain experts, translating business goals into loss [technical] metrics. They formalize data science prototype workflows (but they do not necessarily build them). They can define and dimension the data collection effort and estimate its costs. They can also estimate the effort needed to build and tune the workflow and to put it into production.
In a nutshell, Data Science Formalizers possess all the elements to make informed decisions about building a data-centric product.
Of course, there are a lot of details that can be added for each point mentioned above but the point was to start a conversation. What do you think? Is there something that needs to be added? Any stories to share?
References:
Andrew Ng — AI Lecture at Stanford https://www.youtube.com/watch?v=21EiKfQYZXc
https://medium.com/towards-data-science/the-data-science-ecosystem-industrial-edition-938582427466
Work Rules!: Insights from Inside Google That Will Transform How You Live and Lead by Laszlo Bock
